AI Access & Inclusion Lab
We examine barriers across the AI ecosystem—from computing and network infrastructure to cost, availability, platforms and interfaces, accessibility, and effective use—with particular attention to education, developing regions, and communities underserved by existing AI systems. Taking a civic approach, we also examine how AI is introduced into communities and institutions, how it shapes participation and relationships, and how affected communities can help shape its design and use.
Participating Faculty: Fahad Dogar, Jaylin Hershkovitz, Milo Koretsky, Hari Sundar
Our projects examine different layers of meaningful AI access:
Infrastructure → Cost → Platforms & Interfaces → Effective Institutional Use → Inclusive Design & Evaluation
Projects
Infrastructure Access
Cloud Digital Divide
As AI applications become increasingly compute- and communication-intensive, access to appropriate computing and network infrastructure can determine who receives good performance and who does not. This research studies how the geographic distribution of cloud infrastructure creates disparities in application performance and develops systems approaches for reducing them. Our Divided at the Edge work, for example, examines geographic and socioeconomic disparities in access to low-latency cloud-edge infrastructure.
Cost of AI Access
LLMProxy — Reducing the Cost of AI Access
The cost of using advanced AI models can itself be a barrier to access, particularly as applications make increasing numbers of model calls and use more capable models. One strand of LLMProxy develops mechanisms for reducing these costs while navigating tradeoffs among cost, response quality, and inference time. The system supports cost-saving mechanisms including model selection, context management, and caching, providing a foundation for studying how applications can make advanced AI capabilities more affordable.
Paper: LLMProxy: Reducing Cost to Access Large Language Models
Platforms & Interfaces
LLMProxy — Making AI Easier to Build With
Technical complexity presents a different barrier to AI access. Students and other developers increasingly need to navigate multiple model providers, APIs, and rapidly changing capabilities simply to experiment with AI. LLMProxy provides a common interface across multiple models and common functionality for building AI applications. At Tufts, we have used this infrastructure to give students access to multiple models while reducing the technical friction involved in experimenting with and building AI applications. This work treats access as including not only the ability to use AI, but also the ability to build with it.
Paper: LLMProxy: Reducing Cost to Access Large Language Models
WaLLM — Reaching Users Through Familiar Platforms
Even affordable and technically accessible AI may not reach many people if it requires them to adopt unfamiliar applications or interfaces. WaLLM explores delivering generative AI through WhatsApp, a platform already embedded in everyday communication in many parts of the world. Long-running deployments involving users in the United States, Pakistan, and Somalia provide a setting for studying how familiar interfaces, engagement mechanisms, local context, and trust affect actual AI use. More broadly, this work asks how we can bring AI to people through technologies they already use rather than requiring them to adopt an entirely new AI ecosystem.
Paper: WaLLM: Insights from an LLM-Powered Chatbot Deployment via WhatsApp
Effective Institutional Use
AI in Education
Providing students and institutions with access to AI does not by itself ensure that they can use it effectively. This research develops and studies AI systems embedded within educational environments, including course assistants, academic advising systems, and proactive AI mentors. The work examines how AI interacts with existing institutional processes and human roles, with questions around help-seeking, trust, participation, evaluation, and appropriate human oversight. AdvisingWise, for example, explores a human-in-the-loop approach to academic advising in which AI assists with information retrieval and response generation while academic advisors retain responsibility for validating guidance provided to students.
Inclusive Design & Evaluation
AI for Neurodiversity
AI systems increasingly mediate communication, but they are often designed and evaluated around dominant communication norms. This research examines how AI can support communication across neurotypes and how well current AI systems understand and respond to diverse communication styles. NeuroBridge takes a reciprocal approach to cross-neurotype communication: rather than asking autistic people alone to adapt to neurotypical communication norms, it explores how generative AI can help neurotypical people better understand autistic communication. NeuroBridge received the ACM ASSETS 2025 Best Student Paper Award.
A complementary research direction focuses on evaluation and benchmarking. We are evaluating current AI models across neurodiverse communication styles and developing a broader benchmark for this setting. The goal is to identify systematic strengths and failures in current models and develop evaluation methods that can guide the design of more inclusive AI systems.